Hand Gemini a 90-minute lecture, ask a single question about one slide, and up to now the bill covered every last second of that recording. One frame per second, beginning to end, regardless of whether the answer sat at minute three or minute 83.
That’s what Google is now rewriting. Agent-based video analysis is arriving in several Gemini models, and rather than marching through footage frame by frame at a fixed rate, the model goes looking for the relevant stretches by itself. According to Google, the savings in tokens and cost are substantial.
The model chooses where to look
In the agent-based variant, Google says the model’s reasoning is wired straight into native video tools. Which sections get examined, at what speed, and through which modality — frames, audio or transcript — is the model’s call. Only the moments and signals a given task genuinely requires get pulled in.
An internal tool now fetches nothing more than the relevant slice of the video file. This kind of selective handling was already possible for developers willing to build it by hand. What’s changed is that the plumbing no longer has to be written by anyone.
Sub-second moments that used to disappear
Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite — the newest models — are able to register moments lasting less than a second. Think state changes or cuts, the sort of thing that falls between the cracks when sampling happens once per second. Google’s claim is that this makes automated video editing considerably more precise.
That detail deserves attention. A cut, a flicker, a light switching on — under one-frame-per-second sampling, such events either happen to fall inside a captured frame or, as far as the system is concerned, never happened. Individual scenes buried in hours of footage can also be located without chewing through millions of tokens.
Anomalies get flagged by re-sampling suspicious time windows at a higher frame rate. Repeated movements and individual objects are counted accurately across time as well — precisely the sort of job static sampling has always handled badly.
The real price of static processing
Gemini has leaned on static processing until now: video sampled at a fixed rate, one frame per second by default, adjustable via the API. Ever since native video analysis arrived in 2025, the model has transcribed the audio track and worked through frames on a per-second basis.
Long videos are where the efficiency gains land hardest — 10-minute tutorials, 90-minute lectures, multi-hour recordings alike. Static processing forced developers into a choice between steep token bills and shortcuts that discarded important detail. Neither was a good outcome. In practice, most went cheap and accepted the blind spots.
Across Google’s own benchmarks, LongVideoBench among them, Gemini 3.7 Flash running agent-based analysis posts the highest overall quality along with the best balance of accuracy and cost efficiency. Worth keeping in mind: those are Google’s figures, measured on Google’s benchmarks.
The groundwork was already laid
Underpinning all this is “agentic vision,” which Google rolled out for Gemini 3 Flash in January. That capability had the model writing and executing Python code to zoom, crop and annotate images, reviewing each result in a think-act-observe loop before it answered.
At launch it wasn’t kicking in automatically in every case. Still, the foundation was in place, and when Gemini 3 Flash was announced back in December, Google had already signposted visual and spatial reasoning for video as something still to come. That promise has now landed.
Switching it on, and the price tag
Video uploads and YouTube videos can both use the feature today via the Gemini API in Google AI Studio, as well as on the Gemini Enterprise Agent Platform. Developers flip the processing mode to “agentic” in the API config; billing stays at standard Gemini API token rates with nothing extra tacked on. The Developer Guide carries the fuller details.
There’s no premium tier here and no surcharge attached to the smarter sampling — unusual enough to be worth flagging.
The apps are next
Google intends to push these improvements into its own products as well. A rollout to every Gemini app user on Flash and Flash Lite devices is expected shortly.
In the months ahead, agent-based video analysis is also set to drive the “Ask YouTube” feature on the playback page, so responses hew more closely to what is actually on screen. Anyone who has ever quizzed a chatbot about a YouTube clip and received an answer plainly stitched together from the transcript alone knows the gap this is meant to close.
Already pushing long-form video through the API? Switch the mode to “agentic” and re-run a job you have existing cost data for. That side-by-side will tell you far more than any benchmark chart.